Open redknightlois opened 5 years ago
Thanks for defining this class.. Can you share an example how to use this trainer class along with DDPG and SAC?
Standard examples show how to do that. There is no difference between the current and this one. I use #52 for dataset size reasons though, but for the rest is pretty straightforward.
Thanks.. @redknightlois , do you have a sample replay_buffer compatable with pytorch dataset class? Is env_replay_buffer or any other class in rlkit.data_management is compatable?
Thanks, Narasimha
Hmmm, so it looks like the main difference is the addition of expert_data_collector
. Is that correct? In that case, I'm not sure if we need to create an entirely new class for this. One option would be to add that data to the replay buffer before passing the replay buffer to the algorithm. What do you think of that? It would help separate out the algorithm from the pretraining phase.
This example dataset based trainer also does expert signal recollection, so that is why I didnt do a PR, will let it to you to decide which parts make sense for rlkit.